Skip to main content

telethon-floodgate

Flood-wait handling for Telethon user accounts: a proactive rate-limit gate, a reactive circuit breaker, and flood-wait sleep/retry helpers — three layers that keep a working account out of Telegram's multi-hour FLOOD_WAIT bans.

Telegram does not publish rate limits. This library is the production flood stack of a multi-account Telegram content collector, calibrated against real incidents: a fresh session firing 160+ auth.resolveUsername calls in seconds escalated into 15–18 hour bans, and repeated getDialogs sweeps ran a account into a 14.8-hour ban. The defaults are deliberately conservative and every bucket is configurable.

Install

pip install telethon-floodgate

Requires Python 3.11+, Telethon 1.x and pydantic 2.x.

The three layers

1. Proactive rate-limit gate — before the call

from telethon_floodgate import TelegramRateLimitGate, TelegramRateLimitedError

gate = TelegramRateLimitGate()          # create once, next to your client pool

def before_telegram_call(phone: str, operation: str) -> None:
    category = gate.category_for(operation)   # "send", "history", "dialogs", ...
    retry_after = gate.try_acquire(phone, category)
    if retry_after > 0:
        raise TelegramRateLimitedError(phone, category, retry_after)

Buckets are independent sliding windows keyed by (phone, category): dialogs (1/min — the #1330 incident), dialog_sweep, history, admin_action, send (30/min), channel_lifecycle (3 per 5 min), plus a permissive default. Override any of them:

from telethon_floodgate import RateLimitSpec, TelegramRateLimitGate

gate = TelegramRateLimitGate(
    category_limits={"send": RateLimitSpec(max_calls=10, window_sec=60.0)}
)

An operation tag that is not registered resolves to the permissive default bucket with a warn-once log. Integrations that register their tags can fail closed instead: TelegramRateLimitGate(strict_categories=True) raises TelegramUnregisteredOperationError (a ValueError, deliberately not a refusal) from category_for for an unknown tag — naming the tag, the registered categories/tags and how to register it — while try_acquire keeps its never-raises contract (#29).

resolve and reaction are intentionally no-op categories: in the origin project those paths have dedicated limiters (see ResolveRateLimiter, also shipped here).

Async callers and paginated history

await gate.acquire(phone, category) waits for one category slot, rechecking after each deferral. Cancellation propagates normally. Reserve inside the retry factory so every attempted call is counted:

from telethon_floodgate import run_with_flood_wait_retry

async def send_attempt():
    await gate.acquire(phone, "send")
    return await client.send_message(peer, text)

await run_with_flood_wait_retry(send_attempt, operation="send")

async def fetch_history():
    iterator = gate.wrap_messages_iterator(client.iter_messages(peer, limit=250), phone)
    return [message async for message in iterator]

await run_with_flood_wait_retry(fetch_history, operation="history")

wrap_messages_iterator reserves a history slot for each message-fetch RPC, including search and ID batches, rather than once for the entire iterator. Do not also acquire an outer history slot. The per-iterator proxy leaves the shared client unchanged and passes unrelated requests through. It requires a fresh Telethon-style iterator with a writable, callable client; unsupported or already wrapped iterators fail explicitly. Recreate it inside the retry factory.

gate.acquire and gate.wrap_messages_iterator accept sleep= — pair it with the gate's time_func= in offline tests; the run_with_flood_wait* helpers always use asyncio.sleep. None of these enable per-peer limiting: use try_acquire for per-peer or compound reservations.

TokenBucket(rate_per_min, burst=1, clock=..., sleep=...) provides a separate smooth-refill outgoing cap. await bucket.acquire() waits in a fair queue and logs a warning when exhausted; bucket.enabled is false for non-positive rates. The bucket is per client/event loop, not a cross-process account-wide ceiling. Consumers choose its rate/burst and which logical operations consume tokens; the category gate can independently constrain bursts and retries.

Per-peer send limits

Telegram throttles sending per peer, not just per account: roughly one message per second to the same private chat and about twenty per minute into the same group or channel. Pass a peer key to try_acquire and the gate checks a second, independent (phone, category, peer) bucket before the category one:

from telethon_floodgate import peer_key

retry_after = gate.try_acquire(phone, "send", peer=peer_key(entity))
if retry_after > 0:
    raise TelegramPeerRateLimitedError(phone, peer_key(entity), retry_after)
  • A per-peer refusal does not consume the account-wide category slot, so a burst aimed at one peer cannot burn the account budget.
  • peer_key() derives "user:123" / "channel:-100123" / "chat:-456" / "username:durov" / "id:123" from ints, strings, Telethon TL peers, input peers and full entities — purely local attribute access, never a network round-trip.
  • Only send:user, send:channel and send:chat are configured by default; unknown kinds pass through unthrottled. Override or disable via peer_limits (a "send:user" entry with a permissive spec effectively turns it off) and cap memory with peer_max_buckets (LRU, default 4096).
gate = TelegramRateLimitGate(
    peer_limits={"send:user": RateLimitSpec(max_calls=1, window_sec=5.0)},
    peer_max_buckets=4096,
)

TelegramPeerRateLimitedError subclasses TelegramRateLimitedError, so one "operation unavailable, move on" handler covers every layer.

2. Reactive circuit breaker — when Telegram answers anyway

from telethon_floodgate import FloodCircuitBreaker

breaker = FloodCircuitBreaker(threshold=3, cooldown_seconds=300)

breaker.check(operation, phone)            # raises TelegramOperationSuspendedError
                                           # while the pair is suspended
try:
    await client.some_call()
    breaker.record_success(operation, phone)
except telethon.errors.FloodWaitError:
    breaker.record_flood(operation, phone)

After threshold flood waits on the same (operation, phone) the pair is suspended for cooldown_seconds, then exactly one half-open trial call is allowed. TelegramOperationSuspendedError subclasses TelegramRateLimitedError, so a single "operation unavailable, move on" handler covers both layers.

3. Flood-wait helpers — classify, sleep, retry, report

from telethon_floodgate import (
    HandledFloodWaitError, FloodWaitInfo,
    handle_flood_wait, run_with_flood_wait, run_with_flood_wait_retry,
    is_transient_flood_wait_seconds, is_blocking_flood_wait_until,
)

# wraps one awaitable: FloodWaitError -> HandledFloodWaitError (+ pool report)
await run_with_flood_wait(client.get_dialogs(), operation="warm", phone=phone, pool=pool)

# retries transient waits (<=60s) under a total budget (120s default)
await run_with_flood_wait_retry(factory, operation="fetch", phone=phone, pool=pool)

handle_flood_wait reports the wait back to the pool via await pool.report_flood(phone, seconds) so account rotation can skip the flooded account — persist that to your own storage; the library is storage-agnostic.

Pass gate= (any object with note_flood(phone, seconds), e.g. the TelegramRateLimitGate you already hold) to the same helpers and every handled FLOOD_WAIT also feeds the gate's optional adaptive backoff — no manual wiring needed. Don't ALSO call note_flood from a report_flood hook when passing gate=: that double-counts.

RPC dispatch boundary (opt-in)

Attach to a client before connecting. Every application request sent by its main, borrowed or CDN sender is checked, including nested entity resolution, native Message/conversation/event traffic, transfers, batches and retries. Transparent query wrappers are classified by their inner request, including when receive_updates=False. Unknown RPCs use the bounded default category; strict categories reject unknown registrations.

from telethon_floodgate import RpcDispatchBudget, attach_rpc_gate

username_budget = RpcDispatchBudget(
    peek_batch=lambda calls: username_limiter.peek_many(phone, len(calls)),
    commit_batch=lambda calls: username_limiter.record_many(phone, len(calls)),
    name="resolve_username",
    applies=lambda call: call.rpc == "ResolveUsernameRequest",
)
handle = attach_rpc_gate(client, gate, phone, budgets=[username_budget])
# ... use the client, reconnect without resetting shared budgets ...
handle.detach()

Share the underlying gate and consumer limiter objects between clients of the same account. A batch is preflighted as a whole, with repeated category/peer costs aggregated. All checks pass before any budget commits or request is sent. Budget callbacks are event-loop affine: peek_batch must not spend; commit_batch must not refuse, raise, or await. Legacy admissions callbacks remain validation predicates; migrate stateful hooks to budgets so multiple checks compose without phantom charges.

Consumer budgets see category-exempt reads too. None in RpcDispatchPolicy means no category charge, not an exemption from consumer checks. Only known protocol leaf requests (keepalives, connection configuration and authorization transfer) bypass application budgets. Message page defaults leave logical history accounting to the consumer/wrap_messages_iterator; dialogs pages have their existing separate page bucket. Transfers use default, not send.

Budgets default to mode="defer". A process ceiling may opt into mode="wait": the guard returns proxy Futures, waits without blocking the event loop, then rechecks every budget. Set max_batch_size for a finite-capacity wait budget; an impossible batch is refused rather than waiting forever. Optional prepare refreshes external state before admission. Cancelling any unsent batch member cancels the unsent batch. After sender invocation, cancellation follows native Future semantics and does not prove the write was undelivered.

on_decision receives curated admission/dispatch facts; on_result receives an admission and its exception (or None). Flood exceptions carry rpc, rpc_category, rpc_read_only, source and rpc_dispatched before callers observe them, including already-completed Futures. on_flood(seconds, rpc) remains supported. Choose exactly one flood-feedback owner; enabling both flood_feedback=True and consumer handle_flood_wait(..., gate=...) would report the same flood twice. The guard never automatically replays a write.

A consumer that owns durable flood reporting can put its completion Future on exception.flood_reporting and set exception.flood_reported = True after success. The flood helpers await that owner before considering their legacy pool-report fallback. This prevents an outer send helper from reporting an already scoped resolve flood as an account-wide send flood.

gate_wait_max_sec optionally waits for small category/peer deferrals instead of aborting an interactive flow between prompts. It defaults to zero (the existing refusal behavior). Every wakeup rechecks all budgets; larger gate deferrals and consumer defer budgets still refuse. The caller's operation deadline bounds the flow. Waiting occurs before dispatch, never by replaying an earlier write.

Repeated attach raises; detach restores the original sender callables and cancels pending admission waits. Reattach with the same budget objects. Supported Telethon 1.x dispatch Futures are exercised offline at the sender boundary; transport-level MTProto retransmission/ACK machinery is not counted as a new application RPC admission.

Design notes

  • Everything runs on one event loop with in-memory state and no DB. The gate uses plain deques; TokenBucket serializes its waiters with an asyncio.Lock.
  • try_acquire never sleeps and never raises — it returns seconds to defer, so callers decide whether to reschedule, skip or await.
  • Atomic multi-slot reservations (try_acquire(..., slots=n)) are supported for compound operations. A request wider than a bucket's max_calls (e.g. a 2-message album against the 1-slot send:user peer bucket) can never be admitted; instead of raising, try_acquire defers it at that bucket's window cadence, indefinitely (#7) — split such requests so each fits its narrowest bucket.

Quota evidence

Telegram does not publish a complete numeric FLOOD_WAIT table. The gate uses conservative guardrails from the sources below and keeps every value configurable.

Category Configured guardrail Evidence
history (messages.getHistory) 24 requests / 30s A live probe reached FLOOD_WAIT on request 31 after 30 requests; the page size (1 vs 100) did not change the request threshold. Independent empirical notes report about 30 requests / 30s (copy-history-bot-2).
send:user 1 / 1.1s Community guidance is about one message per second to one private peer; the live Saved Messages test passed at about 1s, and the guardrail adds margin.
send:channel, send:chat 16 / 60s Community guidance is about 20 messages per minute in one group/channel (Telegram Limits); the guardrail keeps a margin.
account-wide send and other categories Existing values No repeatable public numeric quota was found for these method families; they remain internal conservative guardrails.

The history value limits each actual GetHistoryRequest. A consumer must call the gate for every page request, not once for an entire logical export. These values describe observed behavior for a method/account shape; they are not Telegram's universal quotas. Telegram's official error reference only defines FLOOD_WAIT_X as exceeding the allowed attempts for a method and its parameters (API errors).

License

MIT — see LICENSE. Unofficial third-party library; not affiliated with the Telethon project.

Live testing

The offline suite (default pytest, CI) never touches Telegram. Live checks are a separate, explicit activity, in four levels:

  • Level 0 — consumer suite (free): every project using this package exercises it through its own real-Telegram test rig.
  • Level 1 — tests_live/ (opt-in): outside testpaths, so the default run never even collects them. Gated by RUN_FLOODGATE_LIVE_TG=1 plus the account environment:
    export REAL_TG_API_ID=... REAL_TG_API_HASH=... REAL_TG_PHONE=... REAL_TG_SESSION=...
    RUN_FLOODGATE_LIVE_TG=1 python -m pytest tests_live -x -v -s
    
    With the gate closed every test skips; with the gate open but the account env missing they fail loudly naming the variables. Behind a VPN/proxy exit that Telegram refuses, set REAL_TG_PROXY to steer the harness through a working route: socks4|socks5|http://[user:pass@]host:port (opt-in extra pip install 'telethon-floodgate[live]'; unset/empty = direct route as before; a bad value fails loudly naming the variable). test_live_peer_keys (read-only) classifies every real get_dialogs() entity and fails on any UNKNOWN kind or malformed key; test_live_send_gate sends 5 messages to Saved Messages through the real gate and proves the 1.1s user-peer pacing engages with no FloodWaitError. Automatic flood sleeping is disabled, so even short waits fail the run. Pacing is measured between send starts, independently of response latency. Successful runs delete their probes; on a flood the test stops without cleanup and prints the probe IDs to delete after the reported wait. A cleanup failure also fails the test.
  • Level 2 — calibration (manual): scripts/calibrate_send_limits.py probes the raw boundary without our gate — fixed interval points, hard budgets (--max-messages, --max-seconds), auto-stop on the first flood (which is the measurement), JSON report with a verdict against SEND_PEER_USER_SPEC:
    RUN_FLOODGATE_LIVE_TG=1 python scripts/calibrate_send_limits.py
    
  • Level 3 — artifact smoke (per release): install the published wheel into a clean venv and run the level-1 read-only test against it.

Getting a session string from a tg_content_factory installation:

python -m src.main account export-session --phone +7... [--json]

The session string grants full access to the account — pass it via the environment only. Safety rules: use a disposable account, send only to Saved Messages ("me"), respect the budgets, and expect every live run to be a conscious decision. pytest does not read .env itself; use set -a; source .env; set +a if you keep the variables there.

Metadata

Release files for telethon-floodgate 0.1.10

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for telethon-floodgate 0.1.10
File Size Uploaded Provenance Actions
telethon_floodgate-0.1.10.tar.gz 105.5 kB Without provenance Details for telethon_floodgate-0.1.10.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for telethon-floodgate 0.1.10
File Interpreter ABI Platform Actions
telethon_floodgate-0.1.10-py3-none-any.whl Python 3 none any Details for telethon_floodgate-0.1.10-py3-none-any.whl

Total release size: 152.3 kB

Release files / telethon_floodgate-0.1.10.tar.gz

Download URL telethon_floodgate-0.1.10.tar.gz
Size 105.5 kB
Tags Source
SHA-256 checksum
How to use checksums
61244122ded55f0a479f930e15f3f4dadc3eeb9af1880a82c9ab874b2173ad07
BLAKE2b-256 checksum
How to use checksums
e6f536876d633012d4cdd47b3c7c8d492d681f26c9da85239bb34f4cebc0ce70
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.15

Release files / telethon_floodgate-0.1.10-py3-none-any.whl

Download URL telethon_floodgate-0.1.10-py3-none-any.whl
Size 46.8 kB
Tags Python 3
SHA-256 checksum
How to use checksums
de554256f5913cd65d18c7dc095b2f875946361ca390519a40d01aff0ea2df12
BLAKE2b-256 checksum
How to use checksums
db635a427d20d0ce8264a8bdd7a63c8886d9acd2ae7dd7695799b3b75ef3f178
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.15

Security for telethon-floodgate 0.1.10

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page